diff --git a/docker/Dockerfile.base b/docker/Dockerfile.base
index 6be9a57ba..f6cd183bc 100644
--- a/docker/Dockerfile.base
+++ b/docker/Dockerfile.base
@@ -34,6 +34,7 @@ RUN pip install --upgrade pip && \
     pip install --no-cache-dir .${haystack_extras} && \
     pip install --no-cache-dir ./rest_api
 
+
 FROM $base_immage AS final
 
 COPY --from=build-image /opt/venv /opt/venv
@@ -43,7 +44,7 @@ RUN apt-get update && apt-get install -y libfontconfig && rm -rf /var/lib/apt/li
 
 ENV PATH="/opt/venv/bin:$PATH"
 
-# Importing Haystack will generate and persist the json schema, we do this here for two reasons:
+# The JSON schema is lazily generated at first usage, but we do it explicitly here for two reasons:
 # - the schema will be already there when the container runs, saving the generation overhead when a container starts
 # - derived images don't need to write the schema and can run with lower user privileges
-RUN python3 -c "import haystack"
+RUN python3 -c "from haystack.utils.docker import cache_schema; cache_schema()"
diff --git a/haystack/utils/docker.py b/haystack/utils/docker.py
index 5cdc16f9c..a27056845 100644
--- a/haystack/utils/docker.py
+++ b/haystack/utils/docker.py
@@ -1,5 +1,6 @@
 import logging
 from typing import List, Union, Optional
+from haystack.nodes._json_schema import load_schema
 
 
 def cache_models(models: Optional[List[str]] = None, use_auth_token: Optional[Union[str, bool]] = None):
@@ -31,3 +32,15 @@ def cache_models(models: Optional[List[str]] = None, use_auth_token: Optional[Un
         logging.info("Caching %s", model_to_cache)
         transformers.AutoTokenizer.from_pretrained(model_to_cache, use_auth_token=use_auth_token)
         transformers.AutoModel.from_pretrained(model_to_cache, use_auth_token=use_auth_token)
+
+
+def cache_schema():
+    """
+    Generate and persist Haystack JSON schema.
+
+    The schema is lazily generated at first usage, but this might not work in Docker containers
+    when the user running Haystack doesn't have write permissions on the Python installation. By
+    calling this function at Docker image build time, the schema is generated once for all.
+    """
+    # Calling load_schema() will generate the schema as a side effect
+    load_schema()